Assessing Predictive Models for Fairness Based on Activity-Space Patterns

arXiv:2605.23234 · cs.LG, cs.CY · Submitted 2026-05-22 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Assessing Predictive Models for Fairness Based on Activity-Space Patterns".

Jane: The paper was written by the authors from.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: We're kicking things off with a deep dive into "Assessing Predictive Models for Fairness Based on Movement Patterns."

Jane: It's such a heavy-hitting title, Tom, and the authors—Francesco Lettich and his team—are really challenging how we think about bias.

Tom: You mean they're saying our current way of checking for fairness is too limited?

Jane: Exactly, because most studies only look at where a person lives, like a single fixed address.

Tom: So they're suggesting that where you actually spend your time matters more than just your zip code?

Jane: That's the core idea, since your daily routines can reveal so much about your life without you ever saying a word.

Lu: It's a beautiful way to look at the city, seeing it as a web of connections rather than just a collection of static points.

Meng: I see the complexity there, but how do they actually prevent a model from using those connections to discriminate?

Lalam: It's because our movements act as a digital shadow that reflects our cultural and social identity.

Tom: Jane, do you think most people realize their commute or their grocery trips could be used as a proxy for things like income?

Jane: Most people probably don't, which makes the work of Lettich and his colleagues so important for privacy.

Tom: If a model sees you frequently visiting a specific neighborhood, it might incorrectly assume your socioeconomic status.

Lu: It could even start making assumptions about your religion or your family structure based on your stops.

Meng: That sounds like a massive headache for developers who want to build unbiased systems.

Lalam: It's a reminder that our physical presence in a space is a deeply personal form of data.

Tom: We've touched on the title and the big picture, but we need to get into the actual summary of their findings.

Paper discussion segment 2: Jane: We've just talked about how the authors are moving beyond the idea of a single fixed location.

Tom: And they're really pushing into the nuances of how we actually move through those spaces.

Jane: Right, because they identify three specific dimensions that define a movement pattern.

Tom: You mean things like how many different places someone visits?

Jane: Yes, and they also look at how often those visits happen and how long a person stays in one spot.

Lu: That makes so much sense, because a quick stop at a gas station is very different from a long stay at a workplace.

Meng: I'm curious about how they distinguish between a meaningful routine and just a random movement.

Lalam: It's the difference between a passing glance and a place where you actually belong.

Tom: Jane, so they're essentially trying to turn these messy trajectories into something a model can be audited for?

Jane: They are, by trying to see if certain groups of people who share similar movement patterns are treated unfairly.

Tom: That sounds like a way to catch bias that current spatial fairness tools would completely miss.

Lu: It's like they're looking for the hidden rhythms of a community that an algorithm might exploit.

Meng: But how do they actually prove that the model is being unfair rather than just reacting to real data?

Lalam: They're looking for statistical deviations that suggest a systematic penalty for certain lifestyles.

Tom: It's a much more sophisticated way to look at the intersection of geography and social justice.

Jane: We've seen the dimensions they use, so let's see the actual technical heavy lifting they did to make this work.

Paper discussion segment 3: Tom: We've just gone through the dimensions of movement, and now we need to look at the actual methodology.

Jane: They use this really clever process called trajectory segmentation to separate the "stops" from the "moves."

Tom: So they're basically filtering out the noise of someone just driving through a street?

Jane: Precisely, which lets them focus on those meaningful stop segments where people actually spend their time.

Meng: I'm wondering about the computational cost of doing that across thousands of trajectories.

Lu: It's a complex dance of data, but they also have to deal with the "grid problem" to make sure their boundaries don't mess up the results.

Tom: You're talking about the Modifiable Areal Unit Problem, right?

Jane: Yes, so they use multiple different geographic zonings with different resolutions and alignments to stay accurate.

Tom: That sounds like a way to make sure they don't just get lucky with how they draw the lines on a map.

Meng: They also use something called frequent itemset mining to handle all those possible combinations of cells.

Lalam: It's a way of finding the common threads in the way we occupy space.

Tom: And then they hit it with a spatial scan statistic to see if the predicted values are actually skewed.

Jane: They even use Monte Carlo simulations to make sure the results aren't just happening by chance.

Lu: It's a rigorous way to ensure that the digital eyes watching us are actually seeing the truth.

Meng: I'll admit, the way they handle the scale of the search space is pretty impressive for a real-world application.

Tom: With all that math and all those simulations, I wonder how they actually decide if the results are statistically significant.

Conclusion: Tom: We've covered everything from the authors' initial ideas to the heavy technical math behind their approach.

Jane: It's been a lot to take in, especially seeing how they balance the need for detection with the need for precision.

Tom: They found that while it's great for finding unfairness, it can be a bit harder to pinpoint the exact boundaries.

Jane: That trade-off between power and localization is something every developer needs to understand.

Lu: I hope this leads to a future where our cities are designed to be fair by default, not just efficient.

Meng: From a practical side, this gives us a real framework to start auditing these movement-based models right now.

Lalam: It's a step toward ensuring that our digital lives respect the dignity of our physical movements.

Tom: Jane, do you think this is going to change the way regulators look at location data?

Jane: I think it's inevitable, because you can't ignore the bias that's hidden in our daily paths.

Tom: We've reached the end of our look at "Assessing Predictive Models for Fairness Based on Movement Patterns."

Jane: It's a vital piece of research for anyone working at the intersection of AI and society.

Tom: We'll be back with another paper soon, so thanks for listening.

Jane: Goodbye everyone!

cs.LG, cs.CY

Submitted: 2026-05-22

Updated: 2026-09-10

Comments: 35 pages, 10 figures, 7 tables

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 84/100

The gist: This paper introduces a comprehensive methodology for assessing fairness in predictive models by analyzing "activity-space patterns," particularly focusing on movement data.

Key concepts

Activity-Space Patterns: Instead of relying only on a fixed address (like a zip code), this concept suggests that analyzing where a person spends their time—their daily movements—can reveal deep insights into their life, culture, and social identity.
Trajectory Segmentation: This is a methodology used to filter raw movement data. It separates meaningful 'stops' (where people spend time) from the general 'moves' (like driving through a street), allowing researchers to focus on stationary activities.
Modifiable Areal Unit Problem: This refers to the issue that results can be skewed depending on how geographical boundaries are drawn. To maintain accuracy, the authors use multiple different geographic zonings and resolutions in their analysis.

Terminology

Alignment — 11× in this paper · explained in 1 episode
Modifiable Areal Unit Problem — 7× in this paper
Accountability — 4× in this paper · explained in 2 episodes
Activity-Space Patterns — 4× in this paper · explained in 1 episode
Positive Predictive Value — 3× in this paper
Accuracy — 2× in this paper · explained in 1 episode
Core — 2× in this paper · explained in 2 episodes
Aggregation — 1× in this paper · explained in 1 episode
Distribution Shift — 1× in this paper · explained in 6 episodes
Maximum Likelihood Estimator — 1× in this paper
Model-Agnostic — 1× in this paper · explained in 2 episodes
alpha — 1× in this paper · explained in 1 episode

Summary

This paper introduces a comprehensive methodology for assessing fairness in predictive models by analyzing activity-space patterns, particularly focusing on movement data. It provides a rigorous statistical framework to detect spatial and temporal biases—or hotspots of unfairness—that might be embedded within the labeled data, thereby offering critical insights into model reliability and ethical deployment.

Overall Complexity of the Proposed Approach

The assessment process is broken down into five sequential steps: trajectory segmentation (S1), geographic zoning (S2), object-to-cells mapping (S3), identifying subsets for hypothesis testing (S4), and the final hypothesis testing step (S5). The overall complexity is defined as O(S 1 + S 2 + q times (S 3 + S 4) + S 5), where q is the number of geographic zonings used. While time-dominant steps are highlighted as S3, S4, and S5, the authors note that S3 and S4 are embarrassingly parallel with respect to the different zonings, and S5 is embarrassingly parallel with respect to the Monte Carlo simulations.

Hypothesis Testing for Fairness Detection

The core of the method involves formal hypothesis testing using a Bernoulli-based spatial scan statistic. The null hypothesis (H 0) assumes that labels are distributed uniformly, formalized by the likelihood function L 0(theta, PRED). This requires finding the Maximum Likelihood Estimator (MLE) = p(PRED)/n(PRED). The alternative hypothesis (H 1) models the distribution of labels over a specific subset of cells c and the remaining objects, yielding L 1(theta in, theta out, c, PRED). To compute L 1, one must solve for the partial derivatives with respect to theta in and theta out, resulting in MLEs:

in = (p(PRED) - p(c)) / (n(PRED) - n(c))
out = p(c) / n(c)

Computational Cost of Simulation

The computational cost for the hypothesis testing step is highly structured. For each simulation b in 1,, n, the test statistic T must be computed. This involves permuting predicted values (cost O(PRED)) and then computing L(CANDIDATES GZ, PRED b). The cost per subset c in CANDIDATES GZ is O(PRED), leading to a total cost of O(CANDIDATES GZ times PRED) for one simulation. Since L(PRED) can be computed once and shared across simulations, each simulation has cost O(CANDIDATES GZ times PRED).

Interpreting Localization Performance (Sensitivity and PPV)

The paper uses a toy example to clarify the relationship between sensitivity and Positive Predictive Value (PPV). These metrics capture complementary aspects of localization.

High Sensitivity with Low PPV: Indicates that the detected hotspot(s) covers a large portion of the true unfair hotspot(s) but extends substantially into non-unfair areas.
High PPV with Low Sensitivity: Suggests that the detected hotspot(s) is spatially precise, but captures only a limited portion of the true unfair hotspot(s)’s extent.

Assessment Procedure Summary

The final stages of the process follow a standard statistical inference procedure:

Set the desired significance level alpha (Stage 5).
Compute an approximated distribution of the maximum likelihood ratio under H 0 by performing n Monte Carlo simulations (Stage 6).
Compute the maximum likelihood ratio over the original predicted values in PRED, obtaining T obs (Stage 7).
Use the approximated distribution and T obs to determine the Monte Carlo p-value needed to decide whether to reject H 0 (Stage 8).
Store the evidence of movement unfairness (if any) in UNFAIRGZ (Stage 9).

Improvements for AI systems

The provided material describes a highly sophisticated framework for detecting localized anomalies (specifically movement unfairness) using spatial scan statistics and rigorous hypothesis testing on spatio-temporal data. To translate this into next-generation AI systems, we must treat the statistical methodology not as a final output, but as a feature that needs to be integrated into robust, scalable deep learning architectures.

Here are the specific improvements I propose for AI systems:

The current approach relies on sequential segmentation (S 1) and discrete geographic zoning (S 2). This is computationally rigid.

Improvement: Develop a specialized ST-GNN architecture that treats the entire spatio-temporal domain as a dynamic graph G(t).

Nodes: Each moving object's location at time t becomes a node.
Edges: Edges are defined by proximity (spatial constraints) and temporal connectivity (movement trajectory).
Mechanism: The GNN must learn high-dimensional latent representations that capture complex interaction patterns, rather than just simple cell membership. This allows the system to automatically refine zoning boundaries and segment trajectories based on underlying behavioral clusters of unfairness, rather than predefined grids or fixed time windows.

What the improved AI system can do:

Automatic Hypothesis Formulation: It moves beyond predefined zoning (q zones). The system can dynamically propose optimal, non-uniform candidate zones (CANDIDATES GZ) by identifying regions where object interactions violate learned normal behavioral patterns, significantly reducing the reliance on manual grid definition.
Enhanced Feature Richness: It extracts features that are invariant to minor changes in zoning resolution or alignment, making the detection robust against data preprocessing artifacts.

The current bottleneck is iterating over all candidate zones (O(CANDIDATES GZ times PRED)). This brute-force search is computationally expensive.

The reliance on n Monte Carlo simulations (O(PRED times CANDIDATES GZ times n)) is the primary bottleneck for real-time deployment.

The resulting AI system, which I will term the Adaptive Spatio-Temporal Fairness Detector (ASTFD), achieves a paradigm shift from static statistical testing to dynamic, intelligent anomaly detection:

Proactive Zoning: It automatically and dynamically defines the most statistically relevant regions of interest (ROIs) based on learned behavioral deviations, eliminating manual zoning requirements.
Targeted Hypothesis Testing: By using DRL guidance, it maximizes statistical power by focusing computational resources only on the handful of zones most likely to exhibit unfairness.
Real-Time Deployment: By replacing Monte Carlo simulations with VAE surrogates, it achieves sub-second assessment latency, making high-stakes fairness monitoring feasible in mission-critical operational environments (e.g., autonomous vehicle fleets, public transport networks).

In essence, the ASTFD transforms a complex batch statistical analysis into a scalable, predictive machine learning pipeline.

Abstract

Assessing the spatial fairness of predictive models involves establishing whether they are statistically penalizing (favoring) individuals associated with certain geographical locations. Literature on this topic makes the fundamental assumption that each individual is assigned to a single geographical location (e.g., place of residence). However, fairness with respect to the set of regions where one regularly spends time, i.e., the individual's activity space, also matters when fairness is considered. Consequently, we argue that it is necessary to generalize the notion of spatial fairness to also account for such activity-space patterns, leading to the novel problem of assessing predictive models for fairness relative to the movements of individuals. To deal with this problem, we propose an approach that first associates individuals with geographic regions relevant to their activity spaces, considering multiple spatial partitions with different resolutions and alignments, and then employs a suitable spatial scan statistic to assess whether a predictive model is fair based on activity-space patterns. In the experimental evaluation, we study the performance of our approach over thousands of synthetic unfair datasets, showing that it is effective at detecting this new type of unfairness and at retrieving the set of objects treated unfairly, while localization performance exhibits a consistent multi-resolution trade-off.